I am PhD candidate (ABD) in political science at UIUC, my research interests are in the use of computational and statistical methods to study important substantive questions related to mass political behavior and political economy in the American and comparative contexts. I have done research projects on dynamic modelling of mass partisan polarization and segregation, the contextual and temporal variation of individuals’ attitudes towards globalization, model fitting and variable selection of different synthetic control methods, meta-analysis as a binary classification problem, reducing algorithmic bias through more comprehensive model evaluation etc. I am on the 2025-2026 academic job markert.

Publication

Area under the ROC Curve has the Most Consistent Evaluation for Binary Classification PLOS ONE

Selected Working Paper Abstracts

A computational model of the geography of mass partisan polarization (Under Review)

Abstract: Both partisan polarization and partisan segregation are phenomena of great concern in the US in recent decades. Existing research that looks at mass partisan polarization and geography in conjunction focuses exclusively on the extent of partisan geographical sorting but neglects the substantial effect geographical context can have on the formation of mass partisan polarization. This paper shows that while partisan geographical sorting can happen to a certain extent, it only happens in spaces where there is a certain level of partisan segregation to start with. In essence, geographical context’s effect on partisan polarization is more substantial than partisanship’ effect on geographical sorting, a surprising result given current literature’s focus on the latter rather than the former.

The benefits of Lasso for counterfactual prediction with artificial control (Under Review)

Abstract: The synthetic control method is increasingly being used to estimate the effect of an intervention when units of interest are at the aggregate level and there does not exist suitable control units to serve as the counterfactual for the treated unit. However, it can suffer from the problem of unsatisfactory model fit when the artificially constructed synthetic control does not approximate well the pre-intervention outcome trajectory of the treated unit. More specifically, sparsity of the control units that have predictive power for outcome of the treated unit is not sufficiently explored in the literature. Here, we show that a sparsity estimator and Lasso (least absolute shrinkage and selection operator) in particular can not only effectively select control units that are important predictors of the outcome trajectory of the treated unit but also consistently estimate weights for the control units and cross-sectional covariates across different data samples. We showcase these points through simulated data and two empirical examples: effect of Right-to-work legislation on occupational fatal injury rate in the US and economic effect of Hong Kong’s political and economic integration with mainland China.

Fairness of Model Evaluation: How Evaluation Metrics Can Exert Aggregation Bias (Under Review)

Abstract: This paper examines a critical question in application of machine learning models in political science, namely how model evaluation can exert aggregation bias through the use of evaluation metrics. More specifically, it investigates how model evaluation metrics including true positive rate, true negative rate, positive predictive value, negative predictive value, accuracy and Area Under the ROC (receiver operating characteristic) Curve, and F1 score can favor model performance on one class of data over the other class of data for binary classification tasks. Through two political science case studies and a simulation study, our analysis results show that Matthew’s correlation coefficient is the only metric that has high values only when model predicts well on both positive and negative classes of the data while accuracy, F1 score and Area under the ROC Curve all can have high values when the model only predicts well on either the positive or the negative class. These patterns hold under a spectrum of data scenarios in terms of prevalence and for six commonly used machine learning models. The results have significant implications for application of machine learning models in the social and behavior sciences.